How to Track ChatGPT and Perplexity Traffic
AI assistants are sending real traffic and most of it hides in your direct bucket. How to see ChatGPT, Perplexity and Claude referrals in GA4.
The channel that appeared while nobody was tagging
Sometime in the last two years, a new referral channel showed up in everyone's analytics, mostly unlabeled. People ask ChatGPT for tool recommendations, Perplexity for comparisons, Claude for how-tos, and then they click through. Surveys of marketing teams keep finding the same embarrassing gap: the overwhelming majority of brands can't attribute their AI-assistant traffic at all. It's growing, it converts well (these visitors arrive pre-sold by a recommendation), and it's landing in your reports as a mix of referral crumbs and direct-traffic mystery.
The good news: some of it's trackable today with 20 minutes of GA4 configuration. The honest news: a large share is structurally invisible, and you should know which share so you stop under-crediting the channel.
What each assistant actually sends
ChatGPT started appending its own tracking parameter to outbound links in mid-2025: utm_source=chatgpt.com. Clicks from the web app usually also carry a chatgpt.com referrer. That's the trackable part. Clicks from the ChatGPT desktop and mobile apps, though, often ship with no referrer at all, and unless the UTM survived, they land as direct traffic. Estimates of how much AI traffic gets mislabeled this way run from a third to well over half.
Perplexity passes a perplexity.ai referrer from the web reliably, apps less so. Claude links carry a claude.ai referrer when they carry anything. Gemini and AI Overviews in Google Search are the messiest: traffic from AI Overviews is bucketed as plain google / organic, indistinguishable from a classic search click. Google has shown no interest in changing that.
So your AI traffic is currently split three ways: correctly referred (visible), app-launched with no referrer (hiding in direct), and blended into organic search (invisible by design).
The 20-minute GA4 setup
Make the visible part visible. Build a custom channel group:
- Admin, then Data display, then Channel groups. Create a new group (don't edit the default).
- Add a channel called "AI Assistants" with a rule: session source matches the regex
chatgpt|openai|perplexity|claude|copilot|gemini|edgeservices - Put the new channel above Referral and Organic Search in the rule order, so the AI sources get claimed before the generic buckets grab them.
- Use this channel group in your Traffic acquisition reports.
This retroactively reclassifies historical sessions too, since channel groups are computed at report time. The first run is usually an eye-opener: teams find months of AI referrals filed under referral noise.
Worth doing, and worth knowing the ceiling: the regex only catches sessions that arrived with a source. The app-launched clicks aren't in there.
The part only your own links can catch
You can't force ChatGPT's iOS app to send a referrer. What you can control is the URL itself when the URL is yours.
Everywhere your brand feeds the machines, links you publish should be tagged, because parameters survive when referrers don't. That means the URLs in your docs, your llms.txt, your directory listings, your comparison pages, anywhere an assistant might quote a link from. When an AI answer passes along a link that already carries utm_source, the attribution rides inside the URL through any app, any missing referrer, any dark channel.
Watch the harvest side too: check your logged clicks for arrivals carrying utm_source=chatgpt.com. That parameter is ChatGPT tagging itself, and click-level logging sees it even when the session never fires an analytics event.
And mind your utm_medium mapping: pick one medium for AI-facing links (we use ai) and add it to your custom channel rules, or your carefully tagged AI traffic lands in Unassigned, which would be a very 2026 way to lose the data twice.
Measure the mentions, not just the clicks
Last layer: AI traffic starts with being recommended, and clicks undercount recommendations (plenty of users read the answer and never click). Two cheap instruments:
- Ask the assistants yourself, monthly. A dozen buying-intent prompts in your category ("best UTM tool for agencies"), logged in a spreadsheet: are you named, how, versus who?
- Add "AI assistant (ChatGPT, etc.)" to your how-did-you-hear field. It catches the read-but-didn't-click crowd, and it's currently the fastest-growing answer in that field across every team we've compared notes with.
The channel is real, it compounds, and right now most of your competitors are blind to it. A regex, some tagged links and one survey option puts you ahead of the overwhelming majority who can't attribute it at all. Twenty minutes. Go.